Multi-domain Network Service Placement Optimization Using Curriculum Reinforcement Learning

In this paper, we propose a multi-agent Deep Reinforcement learning (DRL) for the placement of network services in a multi-domain context. The objective is to learn how to optimize resources like CPU and memory and maximize the number of accepted services. We apply two stages of learning: vertical and horizontal learning. In vertical learning, the agents learn how to place the functions inside each domain. In horizontal learning, the master agent learns how to divide the slice or service into sub-chains of Virtual Network Functions (VNFs) and how to choose the domains for the placement of each slice VNF. We apply Proximal Policy Optimization (PPO) with curriculum learning and compare our solution to PPO and the Greedy algorithm.